41106f22c0
- Per-file loop count (0-4) with stepper UI on each waveform cell; dynamic MLP output sizing (8 range + N_loops × 4 position outputs) - Beat-synced loop length: global bars/beats selector (2/4/8/16/32/64) with per-loop overrides via long-press dropdown, persistent badges - Global param range groups: dual-thumb range sliders (volume, pitch, position, filter) with ML-driven min/max and per-loop position mapping - Pitch lock toggle (1×) to freeze playback rate across all submodes - Fix: _resizeOutputs now syncs _loopOutputs to prevent stale reads - Fix: granular visuals consolidated per-clip to prevent multi-loop flicker |
||
|---|---|---|
| .. | ||
| c15 | ||
| css | ||
| devlog | ||
| faust | ||
| js | ||
| osc-bridge | ||
| wasm | ||
| a-immersive.html | ||
| b-workbench.html | ||
| c-journey.html | ||
| designs.html | ||
| index.html | ||
| README.md | ||
| serve-coop.py | ||
| serve.sh | ||
| SPEC-controls.md | ||
| SPEC-shapeseq.md | ||
| test-midi-receiver.py | ||
| TODOS.md | ||
NISPS Playground
Browser-based interactive demo of the NISPS ML engine.
Run locally
python3 -m http.server
# Open http://localhost:8000
Run the command from the playground/ directory.
What it does
- Maps 2D joystick input (X/Y) through an MLP (
[3, 32, 48, 64, 126]) to 126 outputs. - Visual mode: first 20 outputs drive a flow-field particle visualization on Canvas2D.
- Synth mode (C15): all 126 outputs control the C15 WASM synthesizer — envelopes, oscillators, shapers, filters, feedback/output mixers, cabinet, and effects.
- Supports two learning modes:
- Examples: add explicit input/output pairs and train.
- RL Feedback: give thumbs up/down while exploring outputs.
C15 synth parameters
The 126 synth parameters (js/synth/param-map.js) cover all sonically meaningful continuous parameters of the C15 engine. Excluded from the C15's 287 total params:
- Hardware routing (56) — no physical MIDI in browser
- Macro controls (12) — conflicts with direct ML control
- Scale/tuning (13) — would break pitch
- Key tracking / velocity (22) — depend on note context ML can't observe
- Envelope mod depths (19) — too many multiplicative interactions for 2-input ML
- Discrete/structural/dangerous (19) — pitch sweep, volume, switches, resets
- Secondary config (15) — curves, chirp, shaper blend, source selects
UI controls
- Output mode tabs (floating bar): switch between Visual and Synth modes.
- Expand / Collapse chevron: expands bottom sheet with training controls, synth settings, advanced param sliders.
- Synth presets: tiered presets (Beginner → Expert) control which of the 126 params the ML engine can modify.
- Help (
?) overlay: in-app usage guide. - Follow mode: double-click joystick in RL mode to toggle no-hold interaction.
- Keyboard in Follow mode (RL):
2= thumbs up,1= thumbs down. - Gamepad in RL:
RB= thumbs up,LB= thumbs down,A= train,X= randomize,B= clear.
Control Surface
The immersive app has a control surface system (Phase 1 of SPEC-controls.md) for tuning how exploration and learning feel.
Compound axes on the floating bar (3 sliders: Bold, Mem, Prec):
| Axis | Controls | Low end | High end |
|---|---|---|---|
| Boldness | Input zoom, noise cap, noise growth, LR, weight decay | Cautious: small changes, heavy regularisation | Bold: full range, explosive exploration, fast learning |
| Memory | Max examples, example decay, noise decay, convergence | Amnesia: only last few interactions matter | Elephant: every example sacred, stable mapping |
| Precision | Input curve, deadzone, smoothing, slew rate, momentum | Raw: 1:1 with physical movement, twitchy | Precise: heavy shaping, deadzones, smooth |
Settings drawer (gear icon, bottom-right): individual param overrides for Input, Training, Exploration, and Output sections. Manual overrides persist as offsets when compound axes move (trim-pot model). Double-tap an axis to re-link all params.
Control presets: Default, First Touch, Jazz Hands, Sculptor, Improviser, Microscope.
Input pipeline: joystick input is processed through deadzone → zoom → curve → smoothing → momentum-as-zoom before reaching the MLP. Zoom narrows the effective input window around an anchor. Zoom-at-zero freezes input.
Enhanced joy-map: zoom minimap with adaptive grid (4×4 → 32×32), vanishing trail with tap-to-return, dual noise rings (zoom + noise level).
Files
Original app (index.html)
index.html- page structure.css/style.css- layout and visual styling.js/app.js- app wiring and interaction logic.
Immersive app (a-immersive.html)
a-immersive.html- fullscreen immersive UI.css/a-immersive.css- immersive layout and styling.js/a-app.js- immersive app wiring, state management, persistence.
Shared modules
js/nisps/- JavaScript MLP + IML core (also WASM variant).js/ui/visualizer.js- flow-field particle system (Canvas2D).js/ui/joystick.js- virtual joystick component.js/ui/gamepad.js- gamepad input handling.js/ui/hand-tracker.js- MediaPipe hand tracking (14 features).js/ui/input-pipeline.js- input processing pipeline (zoom, deadzone, curve, smoothing, momentum).js/ui/control-surface.js- compound axes, override resolution, control presets.js/ui/control-surface-ui.js- settings drawer and floating bar axis sliders.js/ui/joy-map-enhanced.js- zoom minimap, vanishing trail, dual noise rings.js/synth/- C15 WASM bridge, parameter map, presets, arpeggiator.c15/- C15 engine WASM binary and parameter definitions.SPEC-controls.md- comprehensive control surface spec (4 phases).